daskperiment: Reproducibility for Humans

@Scipy Japan 2019/4/24

Transcript

1. daskperiment Reproducibility for Humans Masaaki Horikoshi @ ARISE analytics
2. Self-introduction • Masaaki Horikoshi • ARISE analytics • Data analytics, etc • A member of core developers: • GitHub: https://github.com/sinhrks
3. Contents &YQFSJNFOU
   .BOBHFNFOU EBTLQFSJNFOU
4. Reproducibility • The closeness of the agreement between the results of measurements of the same measurand carried out with same methodology described in the corresponding scientific evidence. (wikipedia) • Without reproducibility: • Unable to replicate the previous experiment trial. • Fruitless investigations and collation works required.
5. Data Analytics Process https://en.wikipedia.org/wiki/Cross-industry_standard_process_for_data_mining • Lots of trials and errors is necessary especially in initial phase. $3*41%.
6. Data Analytics Pipeline %BUB
   QSFQBSBUJPO .PEFMJOH
   &WBMVBUJPO %BUB • Varies along with the trial and errors • Data dependent • CACE (Change Anything, Change Everything) %FQMPZNFOU 5SBJOFE
   .PEFM
7. Dependencies • Data: Data definitions, periods, target domains… • Logic: Feature engineerings, ML algorithms, hyper- parameters… • Environment: Host, interpreter, packages… &OWJSPONFOU &OWJSPONFOU %BUB
   QSFQBSBUJPO .PEFMJOH
   &WBMVBUJPO %BUB %FQMPZNFOU 5SBJOFE
   .PEFM
8. Reproducibility • The closeness of the agreement between the results of measurements of the same measurand carried out with same methodology described in the corresponding scientific evidence. (wikipedia) &YQFSJNFOU
   EFQFOEFODJFT
   NVTU
   CF
   UPUBMMZ
   NBOBHFE
   UP
   HVBSBOUFF
   SFQSPEVDJCJMJUZ
9. Dependency Management • These trials and errors are “smaller” to fit with version control systems. • Minor data modification, feature engineering, new package trial… • Interactive trials and errors sometimes lead to human errors. • Unintended variable overwritten, modified function is not properly applied…
10. Experiment Management • There are lots of tools for logging but… • How often it is checked? • Unable to aware the lost of reproducibility. • Needs “Management”, in addition to “Logging”. • A tool for humans?
11. daskperiment Reproducibility for Humans https://github.com/sinhrks/daskperiment
12. Usage: Jupyter Notebook • Usable in your interpreter with minimum modification. def prepare_data(a, b): return a + b def calculate_score(s): return 10 / s ex = daskperiment.Experiment(id=‘demo_pj') a = ex.parameter('a') b = ex.parameter('b') @ex def prepare_data(a, b): return a + b @ex.result def calculate_score(s): return 10 / s d = prepare_data(a, b) s = calculate_score(d) 4UBOEBSE EBTLQFSJNFOU d = prepare_data(a, b) calculate_score(d) a = 1 b = 2 ex.set_parameters(a=1, b=2) s.compute() %FpOF
    FYQFSJNFOU 3VO
    FYQFSJNFOU %FpOF
    BOE
    SVO
    FYQFSJNFOU %FpOF
    GVODUJPOT %FpOF
    QBSBNFUFST /BUVSBMMZ
    TQMJUT
    EFpOJUJPO
    BOE
    FYFDVUJPO
13. Basic Functionality • Logging history in every experiment trial. • Notify users if required. ...[WARNING] Experiment step result is changed with the same input: (step: calculate_score, args: (7,), kwargs: {}) ex.set_parameters(a=1, b=2) s.compute() ex.set_parameters(a=1, b=2) s.compute()
14. Internal Expressions • Experiment definitions are expressed as computation graph. EBTLQFSJNFOU ex = daskperiment.Experiment(id=‘demo_pj') a = ex.parameter('a') b = ex.parameter('b') @ex def prepare_data(a, b): return a + b @ex.result def calculate_score(s): return 10 / s d = prepare_data(a, b) s = calculate_score(d) ex.set_parameters(a=1, b=2) s.compute() 1BSBNFUFST
    BSF
    DIBOHFE
    QFS
    USJBMT
15. Demonstrations https://github.com/sinhrks/daskperiment/blob/master/notebook/quickstart.ipynb
16. Functionalities Introduced in Demo • Experiment management • Manage trial result and its (hyper) parameters • Save intermediate results and metrics • Detect code and environment change • Validate function purity • Feed random seed • Dashboard • Command line tool (CLI)
17. Usage: CLI • Same definition can be used for CLI execution. ex = daskperiment.Experiment(id=‘demo_pj') a = ex.parameter('a') b = ex.parameter('b') @ex def prepare_data(a, b): return a + b @ex.result def calculate_score(s): return 10 / s d = prepare_data(a, b) s = calculate_score(d) $-* 1BSBNFUFST
    DBO
    CF
    QSPWJEFE
    WJB
    $-*
    +VQZUFS
    /PUFCPPL ex = daskperiment.Experiment(id=‘demo_pj') a = ex.parameter('a') b = ex.parameter('b') @ex def prepare_data(a, b): return a + b @ex.result def calculate_score(s): return 10 / s d = prepare_data(a, b) s = calculate_score(d) ex.set_parameters(a=1, b=2) s.compute()
18. Note: Data Management • Can cover basic usages, but… • Detect external data source change. • Load persisted data to the other experiment. • Using data versioning tool is recommended. @ex def load_user_data(filename): return pd.read_csv(filename) @ex_modeling def load_data(tria_id): return ex_preprocess.get_persisted(‘prepare_data’, trial_id=trial_id)
19. Internal Overview %BUB &OWJSPONFOU &YQFSJNFOU
    NBOBHFNFOU
    %# $PNQVUBUJPO
    %BTL %BUB
    DPMMFDUJPO
    $PMMFDUPS *OQVU
    
    PVUQVU
    #BDLFOE 7JTVBMJ[BUJPO
    #PBSE daskperiment
20. Dask • A flexible parallel computing library for analytics. • Mainly focused on numeric computations. • Provides data structures which is a subset of NumPy and pandas. Blocked Algorithm 4VN $PODBU Dask DataFrame pandas DataFrame 4VN 4VN
21. (Incomplete) List of OSS uses Dask Airflow • (TFLearn) Deep learning library featuring a higher-level API for TensorFlow. • (Distributed Scheduler) A platform to author, schedule and monitor workflows. • Image Processing SciKit. • N-D labeled arrays and datasets in Python. • Executes end-to-end data science and analytics pipelines entirely on GPUs.
22. Dask Computation Graph • Task dependency analysis • Remove unnecessary task (node) • Merge identical task • Static/dynamic computation order analysis • Graph optimization • Inlining • Assigning dependent tasks to the same worker
23. Experiment to Computation Graph • Dask.Delayed: Dask API to built computation graph from arbitrary functions. • Experiment definitions are expressed as computation graph with Dask.Delayed. • Insert pre/post processing per experiment and trial. • Parameters are provided to computation graph per trial. • Dask executes computation graph in parallel (if possible)
24. Collector • Collect environment info and store to experiment DB via backend. $BUFHPSZ
    %FTDSJQUJPO $PEF $PEF
    DPOUFYU 1MBUGPSN *OGPSNBUJPO
    DPMMFDUFE
    CZ
    lQMBUGPSNz
    NPEVMF $16
    *OGPSNBUJPO
    DPMMFDUFE
    CZ
    lDQVJOGPz
    QBDLBHF 1ZUIPO
    *OUFSQSFUFS 1ZUIPO
    WFSTJPO
    WFOW
    JOTUBMMFE
    QBDLBHFTʜ 1BDLBHF
    EFUBJM /VN1Z
    4DJ1Z
    QBOEBT
    DPOEB
    EFUBJMT (JU (JU
    SFQPTJUPSZ
    XPSLJOH
    CSBODIʜ
25. Backend • Users can specify the location to store experiment info as Backend. • Load previous experiment history if exists. • Following backends are supported. • File • Redis • MongoDB daskperiment.Experiment('redis_uri_backend', backend='redis://localhost:6379/0') daskperiment.Experiment('mongo_uri_backend', backend='mongodb://localhost:27017/test_db')
26. Remainings • Collaboration in a team • Optimize computation graph between trials • Extensibility • Fix Collector and Backend API • Support deep learning frameworks
27. Conclusions • It is difficult to guarantee reproducibility in data analytics. • Needs “Management”, in addition to “Logging”. • Built a tool named “daskperiment” • Track and detect anything unexpected. • Experiment steps are automatically parallelized.